Data Science In Practice

Application 1: Recommender systems 

Also known as recommender engines, these are among the most popular real-life applications of data science. They are an information filtering system subclass, what they do is filter out the noise in options available to users before presenting them with only the ones they actually need or will find useful. An example of such filtering is presenting customers with only products they would like on an e-commerce store or filtering possible match on a dating app. Recommendation systems are superior to simple search algorithms in that they introduce you to products you may not have found on your own based on either a collaborative or content-based filtering approach. Collaborative filtering uses your previous behavior and behaviors of other users they are similar to you to gauge your interests while content-based filtering uses assigned characteristic or discrete attributes to give recommendations. 

A hybrid approach that combines both collaborative and content-based filtering can be used just like what the software company Tendril uses. The company supplies consumer solutions and analytics to energy suppliers with the kinds of energy products that customers would typically consider. Data Analytics Manager at Tendril Mark Gately explains that they use Support Vector Regression models to make predictions on how households consume energy. This enables them to provide personalized in-depth information about their customers to their clients. They then use the information in recommending models that help to match customers with existing and new energy products. 

Application 2: Credit scoring 

Most of us already know what the concept of credit scoring is having applied for loans or credit cards in the past. What we may not know of, however, are the sets of decision management rules that evaluate the likelihood of an applicant paying back their debts. Credit scoring is basically a model, made to determine person’s creditworthiness when applying for a loan, mortgage, or credit card. An example of when credit scoring is used is the way financial companies use machine learning models to reach prospects that may have been passed over when traditional banking institutions are used. The FICO score, which was the first credit credit scoring algorithm and has been in use since 1989 is still widely used today. Most direct and peer to peer lending organizations have been focused on developing newer techniques in recent years. Newer algorithms and and machine learning models will be better at capturing relationships and innovative factors that older traditional scorecards for loans would not capture. Examples of complex factors new models capture are whether friends would endorse the applicant and how they manage their cash flows. 

An example of such a company is a leader in mobile consumer and financial technology that developed a complex machine learning and statistical model to facilitates smarter lending decisions. They adopted innovative technologies and used creative approaches to reinvent how business and customers access loans. This approaches enables them reach more customers than traditional methods allowed. 

Application 3: Dynamic pricing 

Businesses use dynamic pricing algorithms to base rates against the level of supply, exogenous factors like time, competitor pricing, and demand. This model works by using popular techniques of classification trees and generalized linear models to get estimates on the right price that would allow them to make the most revenue although the strategies may vary in some companies.  A good example of dynamic pricing is when a ride sharing company uses this model to suggest prices for the people who use their service. 

Many other fields also use dynamic pricing like airlines and ticket admissions who also use dynamic pricing to maximize their revenue. 

According to the Director of Data Science and Analytics of a peer to peer car rental service that operates in more than 2,500 cities, dynamic pricing helps them balance demand and supply so both hosts and travelers get a good deal. After starting out with modeling supply and demand dynamic so naturally dynamic pricing was the next step. Realizing the gap between model deployment and development was wider than what they expected, they needed a broader spectrum of skills from best practices, software architecture, and knowledge of statistical modeling. 

Application 4: customer churn 

When a customer abandons a product or service without purchasing it after entering into your sales funnel, it is referred to as a churn. The need to understand why customers churn is especially significant for subscription based models like gym memberships, monthly box subscriptions and cable subscriptions. A number of algorithms are considered by data scientists when they seek to understand why customers churn and predict them like random forest, k-nearest-neighbors or support vector machines. Data scientists also have to find a way to balance the trade off between precision and recall of their preferred model in addition to its accuracy. 

So they must decide which is more important, identifying a few churning customers without mislabeling any or finding every single churning customer with the possibility of mislabeling a few non churning customers among them. Making this decision requires years of experience and a deep knowledge of the business.  A real life example of this is EAB, The advisory Board company’s’ education division collecting data from demographics, standardized test scores, transcripts and other sources and combining them to spot students who run the risk of failing to graduate to enable them step in and  help them graduate. 

Application 5: Fraud Detection

FinTech companies usually offer services like investing, payment processing via software instead of traditional methods, and banking. These companies process a huge volume of financial transactions every day so they need a reliable and quantifiable way to detect and prevent fraudulent transactions. It works by using binary classifications to spot the problem of whether a transaction is legitimate or not. When a financial institution has thousands of legitimate transactions everyday compared to the few instances of fraudulent transactions, it can be tricky to use a traditional fraud detection binary classification problem. 

But when even one fraudulent transaction can result in a loss of millions for the company, data scientists use a pairing of anomaly detection algorithms and supervised classification techniques to identify suspicious behavior and pick them out.  VIA SMS Group is a financial company that loans over €60 million to their customers in different countries every year. Before granting a loan application, their risk analytics team uses advanced algorithms to see if a customer is fraudulent before deciding to underwrite a loan or reject it. The team uses R programming language to write their decision algorithms before implementing  them using PHP server side language into their mobile apps and web. This enables them to combine state of the art algorithms and complex data lookups to spot fraudulent activity. 

The data science lifecycle 

From the five applications above, we can see how data is used to optimize processes, improve products and services, and inform decisions in a variety of business challenges. Irrespective of what the question is, a data lifecycle typically begins with choosing the right strategy and then implementing and applying it into situations where the business will realize value.